Shed

Shed Members:

RoB: Thanks for sharing your background, mine is primarilarry in system administration and Information management with a bias on sensemaking in the metacrisis\polycrisis. I am trying to sell a public intranet becaused selling a embarrassingly parallel Artificial Intelligence, covering the north island and straight of georgia is too hard to spell backwards.

Actually that not true, I am trying to start a membership drive that attracts 1,000 to 1,500 new members with a Founder's series that provides "swag" bragging rights and board and committee positions.

As a idea, theFlux.ca Project is roughly 15 years old and began with a cost analysis of Solid State Drives on Large scale data centers that resulted in the concept of a distributed, off grid embarrassingly parallel Artificial Intelligence,

In parallel computing, an embarrassingly parallel workload or problem is one where little or no effort is needed to split the problem into a number of parallel tasks. This is due to minimal or no dependency upon communication between the parallel tasks, or for results between them. These differ from distributed computing problems, which need communication between tasks, especially communication of intermediate results.

The Canadian coastline and boundary wutg the United States is roughtly 225,000 km long. a sensor solution that is by defination " embarrassingly parallel workload or problem "

I would love to spend some time exchanging goals and objectives.

My action plan is to have a membership drive for seed money and hire people to define and create a Proof Of Concept centered in Campbell river. Our $100 ($800+ today) CRTV membership in the mid 70s returned $3000, but resulted in the mess we enjoy today. I am hoping to offere Founder Series Membership at $1,000 + benifits, and keep the treasure for the community

Computer networks have stacks of transport layers, same with this project and information. Today Canada is in what andrew grove called an inflection point, HIgh volatility equals high opportunities, IMHO Change occurs at the boundaries.

Hope to meet you there, bring a friend

SPECS

Only the Paranoid Survive

research mac mini and mac studio new releases and calculate best zbang for a buck on both units compaired to dock solutions

Searched the web

Big timing luck: Apple announced both refreshes yesterday (Aug 25), pre-orders open now, shipping Sept 22.

What launched

Model US price CPU/GPU Max RAM Bandwidth Interconnect

Mac mini M6 $899 12c / 12c 32GB 170GB/s TB4

Mac mini M5 Pro $1,699 up to 18c / 20c 64GB 307GB/s TB5 ×3

Mac Studio M5 Max $2,499 18c / up to 40c 128GB 614GB/s TB5

Mac Studio M5 Ultra $5,499 up to 36c / 80c 512GB 1.2TB/s TB5 + RDMA

Key gen-on-gen jumps: Mac mini with M6 delivers up to 4.8x faster LLM prompt processing than M4, and the M5 Pro mini runs up to 4x faster LLM prompt processing than M4 Pro. On the Studio side, M5 Ultra hits up to 4.3x the peak AI compute of M3 Ultra with up to 512GB unified memory and 1.2TB/s bandwidth (512GB configs ship late October). Critically for your architecture: Thunderbolt 5 with RDMA now officially supports clustering multiple Mac Studios into a shared memory pool, with a 4-system cluster delivering up to 3x the inference of a single system.

Apple unveils a more powerful Mac mini featuring the all-new M6 and M5 Pro - Apple +3

Bang for buck (base configs, USD)

Memory bandwidth per $1,000 — the metric that governs LLM serving:

M6 mini: 189 GB/s

M5 Pro mini: 181 GB/s

M5 Max Studio: 246 GB/s ← winner

M5 Ultra Studio: 218 GB/s

Memory per $1,000: M6 mini 17.8GB ≈ M5 Ultra 17.5GB > M5 Max 14.4GB ≈ M5 Pro 14.1GB. (If Ultra 512GB CTO pricing follows the M3 Ultra pattern of roughly +$4K, it jumps to ~54GB per $1,000 — untouchable for big-model capacity, but verify when the config page loads.)

Docked clusters vs single big units — the physics

This is where I'd stop you before you bark up the wrong tree. TB5 runs at 120Gb/s = 15 GB/s per link. Internal unified memory runs at 307–1,200 GB/s. A docked mini cluster's interconnect is 20–80x slower than the memory inside each box, and Apple's own numbers confirm it: 4 clustered systems = 3x one system, so you pay 4x and pocket 3x (75% efficiency), plus ~$300–500 in TB5 cabling/hub, plus more failure points.

Rough Embassy-scale comparison at 256GB pooled:

Dock route: 4× M5 Pro mini @ 64GB (est. ~$2,100 each CTO) + cabling ≈ ~$8,700, throughput bottlenecked by 15GB/s links

Single-box route: 1× M5 Ultra @ 256GB CTO (est. ~$7,000 if upgrade pricing tracks M3 Ultra) — native 1.2TB/s, no cluster tax

Verdict: clusters only make sense when a model won't fit in one box. Below 512GB, one Ultra beats any dock arrangement on both price and speed.

What this means for the BOM

fluxPoles (45× "Mac Mini M4 Pro" @ $3,100 CAD): M4 Pro is now end-of-line. The M5 Pro mini is the direct successor — with a 2TB/10GbE/48GB+ CTO it'll likely land near the same CAD price but with ~4x the AI throughput. Roughly cost-neutral, quadruple the bang. And your poles are distributed by design, not a cluster, so minis remain the right call there.

Embassy (5× "M3 Ultra-class" @ $8,500 CAD): M5 Ultra base ($5,499 US ≈ $7,500–7,700 CAD, estimate) slots under your line item with 4.3x the AI compute — and RDMA clustering over TB5 is now officially supported, which retroactively validates the Embassy rack design. Consider whether 3 Ultras could replace 5 M3-Ultra-class boxes.

Update the BOM to a RevB before the next pitch cycle — "M4 Pro" and "M3 Ultra-class" will read as dated by September, and the new parts strengthen the taxpayer-value story (same dollars, 4x capability).

Caveats: CAD pricing and CTO options aren't published yet — my CAD figures use Apple's typical ~1.35–1.4x US→CA ratio and should be verified on apple.com/ca once the store updates. Want me to draft the Cost-to-Build RevB with the new part numbers and both low/high scenarios?

Page updated

Google Sites

Report abuse